Skyfall AI announced Steve-001 on November 9. The company bought three small software businesses and let an AI system run them. The Neuron reports that one business increased monthly revenue by 127.7 percent under Steve. Another remained 3 percent below its purchase level.
A world model simulates each move
Steve uses what Skyfall calls a code world model. The model is an executable program built from business records. It describes how customers, product usage, revenue and costs push on one another. Before Steve acts, it simulates candidate moves through the model. Each proposal carries a forecast, a timeframe and a confidence level. The gap between prediction and result feeds the next version. Both the model and the plan get updated.
Two software agents got stuck passing the same task back and forth during a website update. Steve changed its own coordination rules to prevent a repeat. It added an agent that steps in when agents disagree. That agent decides how to resolve the issue. The new agent had to work within the four tiers Skyfall sets.
Four tiers decide what runs alone
Every decision lands in one of four tiers. The action runs automatically, runs and alerts the owner, waits for approval or stops. A refund of thirty dollars or less might go through automatically. Anything larger waits for a person. Production releases and spending above agreed limits need approval. Skyfall says the improvements Steve makes to itself have to stay inside those limits.
The tier that catches an action depends on financial size and customer impact. Whether the move is reversible also matters. A password reset is tier one. A pricing change is tier two. A production deploy is tier three. Exporting customer data is tier four. The system that sorts actions by risk is code the operator writes.
Atelier asks me before its agent sends, posts or drafts an email. OpenAI agents disabled monitoring and one denied it. I would check whether the agent that asks permission today can rewrite the code. Could it change the part that makes it ask? The Coen ERP gates every action with mandatory human review. I would want that gate to apply to any change the agent makes. That includes its own approval rules.
AI Trader has hard caps in code. Every stock entry gets a stop-loss order held at the broker. Autopilot is off by default, so each new order waits for my approval. A policy engine sorts actions by risk, financial cost and customer impact. Steve uses that pattern. The difference is whose money is at stake. How fast you can reverse a bad call also matters.
The businesses are real
Skyfall did not say which three businesses Steve runs. The company said each one generates between five thousand and fifty thousand dollars a month. Steve handles customer support, billing disputes, feature requests and software updates. It decides which feature to build next based on support volume and revenue per customer segment.
One business saw churn drop after Steve identified customers likely to leave and offered targeted discounts. Another saw no change in retention. It cut support response time from four hours to eleven minutes. The third business lost a major customer after Steve raised prices without simulating competitor positioning. The raise stayed within tier two, so it ran and alerted the owner after the fact.
The surface governs autonomy.
Skyfall is offering two million dollars to acquire five more businesses for Steve to run. Somewhere a founder is about to hand over the keys and watch a dashboard.